Causal Cross-View Memory for Online 3D Multi-Object Tracking
Received: 09 Nov 2026 Revised: 10 Nov 2026 Accepted: 11 Nov 2026
Published: 2026, vol. 30, issue 4, pp. 369–372
Abstract
In DeepFusionMOT, Stage-4 updates counters but not the stored 2D memory, which is reused when projection is unavailable or LiDAR misses occur. P3DR projects the predicted 3D box as evidence; C2MC writes the accepted 2D state for later frames. On KITTI Tracking Val11, relative to DeepFusionMOT, HOTA rises from 79.876% to 80.385%, sAMOTA from 93.01% to 95.42%, and FP falls from 294 to 110. In a same-source control (RRC disabled; native or projected boxes from one 3D detector), sequence outputs match the baseline.
Keywords: 3D multi-object tracking, camera--LiDAR fusion, cross-view memory, online association.
BibTeX
@article{IS-Li-Mamonova2026,
author = {Li, Jiuming and Mamonova, Tatiana Egorovna},
title = {{Causal Cross-View Memory for Online 3D Multi-Object Tracking}},
journal = {Intelligent Systems. Theory and Applications},
year = {2026},
volume = {30},
number = {4},
pages = {369--372},
}
AMSBIB
\Bibitem{IS-Li-Mamonova2026}
\by J.~Li, T.\,E.~Mamonova
\paper Causal Cross-View Memory for Online 3D Multi-Object Tracking
\jour Intelligent Systems. Theory and Applications
\yr 2026
\vol 30
\issue 4
\pages 369--372
\lang In Russian
RU